AI LEARNS TENNIS
Вставка
- Опубліковано 9 чер 2024
- A Neural Network trained via reinforcement learning learns how to play Tennis. We explore different ways of nudging the A.I using Ml-Agents. The Algorithm used to train this A.I. is called PPO (developed by OpenAI). My name is Sebastian Schuchmann from Germany and I hope you enjoyed this video.
If you want to support me! :)
Patreon: www.patreon.com/user?u=25285137
Twitter: / sebastianschuc7
If you want to get in contact with me, go to my Webform :)
www.sebastian-schuchmann.com/...
The environment shown in this video is part of the ML-Agents examples made by Unity3D. unity3d.com/de/machine-learning - Наука та технологія
I think it'd be cool to train on tennis with a more competitive scoring system: Lose points for missing the ball, gain points for scoring. With this, limiting the acceleration of the racket would promote more advanced playstyles. The position and speed of the opponent's racket would enable even more advanced playstyles. I'd love to see if they'd bounce the ball tactically.
Seems like this video should have more vides
its a very underrated channel
Great video Sebastian, I like how you clearly define the reward function and state space, makes it very clear what is going on.
Awesome video! You did a great job explaining it.
Interesting discoveries packed in an entertaining video - love it!
You are an underrated youtuber
Could you please make a tutorial of how you created such a project?
Have scalable models for your racket - when the AI gets good enough to reach equilibrium with self-play consistently, automatically decrease the size of the tennis tennis racquet. This is the same idea as when a GAN first trains on low resolution images and increases resolution over the epochs.
Oh, interesting. Thanks for the sharing knowlede
Late to the MLA Party, but have you tried advancing this a bit? Giving the computer the ability to rotate the racket so they don't always hit it at 45 degrees? Allowing for high and low shots? Then you can open up another axis of rotation to allow them to place the ball left and right. I think those steps would train a really competitive tennis AI.
How does this guy only have 351 subs?
Interesting, I like it.
Have you tried adding an extra ball? I have, and the agents made the ball collide like bullets in a movie lol.
Unfortunately, I hadn't modified double hit rules to exclude balls that bounced off each other, so the agents simply avoided the balls, fearing double hit losses, and left their victory at the hands of RNG. Once I didt that, now the agent gets overwhelmed by 2 distant balls at once...
What about adding players? Have you tried that as well?
next MMA or boxing or muay thai
A Capture the flag ai vs. battle
Hey how about a chess ai
Air Hockey